Photovoltaic Power Adaptive Hybrid Forecasting Model Integrated with Multi-Dimensional Error Compensation

Abstract

INTRODUCTION: With the deepening of China's "Dual Carbon" goals, the scale of photovoltaic (PV) installations continues to expand, placing stringent requirements on the accuracy, scenario adaptability, and robustness of day-ahead power prediction for PV plants in grid frequency regulation, peak shaving, and electricity market trading. However, existing single prediction models have obvious limitations: linear models struggle to capture nonlinear power disturbances under cloudy days with sudden irradiance changes and extreme weather; nonlinear models are prone to overfitting during stable, high-irradiance sunny periods, leading to redundant accuracy; and most methods lack sufficient robustness against meteorological fluctuations and data noise, resulting in large prediction errors that seriously affect the economy and security of grid operation.   OBJECTIVES: Aiming at the problems in PV plant prediction where a single model finds it difficult to balance linear patterns and nonlinear disturbances, and prediction accuracy is greatly affected by meteorological fluctuations and data noise, this paper proposes a hybrid prediction model based on Linear Regression and Gradient Boosting Tree, along with a multi-dimensional error compensation mechanism.   METHODS:  1. Otptimize feature engineering design, selecting time features, historical power lag terms, and meteorological interaction features as inputs, unifying their scales via Z-score standardization to simplify model complexity and construct a high-quality training set. 2. Design a dynamic weight allocation strategy based on irradiance intensity grading and intraday time periods, integrating the precise fitting advantage of Linear Regression during high-irradiance periods with the nonlinear fluctuation capture capability of Gradient Boosting Tree to establish a hybrid prediction model. 3. Use hourly operational data from a specific PV plant from May 2024 to March 2025 as a sample, performing validation combining an error compensation mechanism composed of sliding window error correction, extreme weather compensation, and residual feedback.   RESULT: The test set RMSE of the hybrid model decreased by 18.2% and 22.5% compared to the single Linear Regression and Gradient Boosting Tree models respectively, with the error deviation rate under extreme weather controlled within 12%.   CONCLUSION: These results verify the effectiveness and practicality of the proposed hybrid prediction model and error compensation method.